Ara Pulido: So thanks, thanks a lot for coming. My name is Ara Pulido. I am engineering manager at Bitnami. This is actually my second KubeCon so still pretty new to the community. And for the next few minutes I'm going to be talking about a tool that we discover in the previous KubeCon in Austin. And since then has become a key piece so far in our development process.
Ara Pulido: So, first of all, who we are: so, Bitnami is a leader in application packaging. We have more than 120 applications on the cloud marketplaces, like AWS, Azure, DCP, etc. And that's how a lot of people know us. And a couple of years ago we started investing a lot in containers and Kubernetes. With the application packaging in our DNA, we invested a lot in Helm and we actually continue maintaining more than 20 of the official charts there.
Ara Pulido: We also do a bunch of opensource projects that are always on the application deployment area of Kubernetes, so we have Kubeless, which is a server less framework for Kubernetes, Kubeapps which is an application dashboard for Kubernetes for your cluster, which is actually the application that we use telepresence for. And others like SSLSecrets or Kube CFG that we develop with HEPTIO, SSLSecrets, and others.
Ara Pulido: And since we started investing a lot in Kubernetes, we also did that internally, so we run many of our workloads in our Kubernetes cluster, and we love Kubernetes as a production environment for your containerize applications. But it seems like the development process for it, it was a little bit of an afterthought. When you're developing for Kubernetes, suddenly it becomes very difficult to manage. So, just to put an example of a very, very simple cloud native application.
Ara Pulido: Very quickly it becomes this microservice architecture where you may have different ingress rules that redirect traffic to a front end, maybe a web UI, and that web UI may talk to one or different services on the back end. And those services may talk to each other and they may need to to talk to, for example, and external database or another external service that may not be in your cluster. So even with a very, very simple app, it certainly becomes this difficult mess to handle on the development process.
Ara Pulido: So, how we've been doing that is historically is basically replicating our architecture in production, in a docker compose, and making this little tricks that we will do. Like we may have a conflict mapping Kubernetes, and we may replace that with just a volume mount on docker compose. And we may replace our secrets with just a envirable, and we may have several services describe that for service recovery, and we maybe have our volumes, etc, etc. And as the application grows, you have to keep maintaining those two manifests.
Ara Pulido: So the manifests that you ran in production, the ones that you use in docker compose to develop and they may divert and you may find regressions or things do not work, sadly, as you expected when you go to your cluster. Also there are things that are not that easily described on a docker compose like your ingress rules or jobs and cromjobs, or your init containers, or your [inaudible 00:03:51] rules, and your Kubeconfig, etc, etc.
Ara Pulido: So, then there is the second solution with having this development cycle off. You bill your image, you make a change, you bill your image, you push it somewhere to a registry, hopefully on your cluster maybe somewhere else. Change your manifest or maybe your manifest are pointing to a def image directly so you test your changes. That's some work and you do that over and over again. And this can become very, very slow. You can be doing things like making sure that you're building your image as part of your bill process to accelerate things, but again it's always a little bit slow. And even if you have to push somewhere in the register you may have these issues of making it a little bit more slow.
Ara Pulido: So, obviously this is a problem that a lot of people are thinking of and there are several solutions that are there. And as you said, we discovered Telepresence last Kubecon. It's a tool by a company called Datawire and very, very high level. What it does is allows you to run a process that is fully connected to a cluster and that cluster could be on your local environment with Minicube, but it also and ideally, will be elsewhere in the cloud. So you can have all the computer that you need or all the memory that you need to run the different services that you're not actually developing but your application may need to talk to.
Ara Pulido: How it works is Telepresence will deploy a two-way proxy into your cluster that start proxing data from your cluster to our client running or your local machine and back. Basically what it gives you, you get service discovery with full DNS resolution for your local machines. So you can talk to any of the services that are in your cluster in a seemless way. You have volumes, you have environmental variables, so all the things that you may need from your cluster when you're developing a service that you are planning to put on that cluster directly from your local machine. And I think that the gifing and the thing that we like about Telepresence is that it does that without you even almost noticing. So you run a local process and it's talking to everything on your cluster without you having to do any change on the code because of this using the same service discovery volumes, environmental variables, etc.
Ara Pulido: Telepresence has several proxy methods. Each of those has pros and cons. You can use the one that you think is gonna serve you better. The first one is VPN, where it creates basically a SSH tunnel between the proxy running the cluster and the client creating more or less like a VPN. The only downside is you cannot use another VPN on top and actually I was testing the demo the other day and I was freaking because it wasn't working. It was because I was actually running on my company VPN.
Ara Pulido: Then you have Inject-TPC, which basically injects a shared library into the process that you want to run. And with a clear limitation that it doesn't work with statically linked process that go.
Ara Pulido: And then docker which is very similar to VPN process with the difference that instead of running just a process it allows you to run an image that you already built. Which works well if you already have this container native development process where you'll build processes not building your app. It's building your app and building your a docker image.
Ara Pulido: So the idea of this talk more than talking about is talking at least to show you how it feels using Telepresence, so I prepare a couple of demos. And let's see came to explain the first one.
Ara Pulido: Is this big enough? Or maybe bigger?
Ara Pulido: Better?
Speaker 2: Yeah.
Ara Pulido: Okay. So the first one is very simple. It's just to show you all these service discovery in volumes, etc., that you can have. So we have these couple of pods in cluster but when I'm in a Kube just to avoid using networking too much. Yes, the idea is that you would have development cluster in the cloud or in your service and primaries to do this. So yeah, we have this service, a simple service called QRTM. The only thing it does is listening on port 5000. When you make a request it remains adjacent with a random quote. It's just as simple as that. We use it as an example of a service that may be running on your cluster, not the one that you're developing that, but you need to talk to it as far off your application.
Ara Pulido: And then we have these Telepresence test that- let's check what it is. [inaudible 00:10:27] deployment. It has an init container so you can see that we can still use init containers. The only thing that it does is mounting the volume call datum into a mount [inaudible 00:10:45] data and just creates just a simple random stream on a file. And then we have our main container which has here an environmental variable. It's not a real secret in this case, but it could be a key ref to a real secret. It will still work. And tFhen you have the same volume mount and basically it reads the file it was supposed to be created by in the container. And the volume is just simple empty there.
Ara Pulido: So obviously if we check the logs of that pod, we have random data there as we expected. And now let's remind you that there is something not going okay. We didn't get random data there, there's something going wrong. We may thing that it's something with the volumes. Something with the init container, want to debug a little bit.
Ara Pulido: So basically we are going to tell the presence to stop the deployment call to the presence test and let's watch here the ports as well. But, I'm not telling the presence to run any process on top, so what it's going to do is to start a bash session that is already connected to my cluster. So as you can see, what is happening is it's swapping the deployment and we will see when it finished what it's actually doing. So and the obviously swapping is changing the deployment and the deployment controller, the replica controller is doing it's thing. And basically some things that it says here, so it says that it's starting the VPN method. That's the full one, so that's the one it's going to use if you don't tell it otherwise. And then it says volumes are routed to the Telepresence route and it says that no traffic is being forwarded from the deployment to my pod. We will see in the second demo how to do that.
Ara Pulido: So let's check. So one of the pods is terminating the other one is starting, so let's check how these pod look like. So you can see we have the volumes now, we have our init containers still here, and we have the main container with our environmental variable. Some special variables coming from Telepresence. As you can see the image has changed from the bc box image that I was using to the Telepresence proxy image.
Ara Pulido: So back to the shell that is connected to my classes. Let's check first volumes, so volumes are routed at Telepresence route so let's go there. So we can see that the data volume was correctly mounted and if we check the data there our init container created that file correctly. So we can see that way we can have access to the volume and see if something was wrong there. Also, we have several environmental variables, some of them are coming from my local machine. But also, I have access to the ones that are on my pod. Here's the one that was on the pod description, and it's still there so I can have access to any of the environmental variables there.
Ara Pulido: Lastly, remember that we were running the service on port 5000 called QOTM. That basically returned adjacent with a random code. So directly from here, from my local machine, I can call QOTM on and I can use the same DNS: service discovery [inaudible 00:15:45]. So if it's running on a different name it's best that I have access to I could use dot [inaudible 00:15:50] etc, etc, to get access to it. And as you can see it works. It feels like I'm [inaudible 00:16:01] into a container inside my cluster, but actually I'm not. As you can see it's still running so if I go here it's my local machine. If I exit here, it's going to the opposite. It's going to replace back the container by the original one and leaving my cluster as it was before I start Telepresence.
Ara Pulido: So this is a very simple demo but it already shows what you can do with Telepresence. I'm going to show now how we're using it to develop Kubeapps, which is an application dashboard for your Kubernetes cluster. Kubeapps comes with the CLI, I'm gonna make that bigger. So it comes with the CLI tool. If you do Kubeapps app it will deploy all the different components that are part of Kubeapps into your cluster and you can see that there are different deployments, also one in the Kubeless name space.
Ara Pulido: The one that we are going to use Telepresence for is Kubeapps dashboard UI. What is this UI? So if we run Kubeapps dashboard, it's going to start this application dashboard. What you can do with, you have access to your helm deployments, we are using actually not only using Tele directly but using something we call Helm CRD. Which for every deployment creates a CRD object and there is a controller looking after it. We have a set of charts available, some different repos, you can add repositories if you need to. This thing for example is happening also with a different main, and different CRD.
Ara Pulido: Functions through Kubeless, we have function as a service in Kubeapps with Kubeless integration. So you can have any function that you may run here. Each of these functions in Kubeless is again, a CRD. Call function. It has service catalog integration. Service catalog is an API extension for Kubernetes that allows you to connect external services provided by what is called the service broker into your cluster. And again, It's an extension of the API. So you can see, this UI has a lot of knowledge about your Kubernetes cluster and it would be very, very difficult to just try and to develop this UI locally.
Ara Pulido: So what we are going to do is we are going to tell Telepresence to swap the deployment call Kubeapps dashboard UI, which is the UI we just saw, on the nameless space Kubeapps because it's not running on the default nameless space. We using in this case the Net TCP method, although there are people on our developing team that are used to VPN as well so both work well. Also we are telling it to redirect all the traffic to on the cluster on port 8080 on that pod to port 3000 on the local port. And the reason why we are doing that is because we are then going to run a development server for that application that is running on port 3000 by default.
Ara Pulido: So again, it's starting the proxy this time with a Net TCP, it's warning you about some limitations of that. Same thing, volumes are routed to a Telepresence route and it's forwarding traffic from port 8080 to 3000. So let's run a development server on port 3000. It's going to take some seconds to start. So it says now that it's listening on local host 3000, but instead we are going to access through this proxy to our cluster that we just created. And actually now that it refreshes, this UI is actually running on my local machine through port 3000. And I still have access to all the service catalog classes that I have for this particular broker, functions, the charts, applications, etc, etc. To demonstrate that's true, we are going to do a quick live change.
Ara Pulido: So this particular file is rendering the view here, the applications. So this Heather, H1 Heather, is that same Heather that we saw. And we are just going to save that and our development server is going realize that there has been a change and it's going to reload the page and it's showing the change. So that way allows us to do a lot faster development process for this particular application, and since we've started using Telepresence it has been a lot easier to make these changes. When making sure that we were able to keep those services running on that cluster talking to our local servers. So that's it for the demo.
Ara Pulido: The only thing left for this talk, I'm going to ask Richard Lee, who is the CEO of Datawire. Thanks for coming, Richard.
Richard Lee: Hi.
Ara Pulido: Applause for him.
Ara Pulido: And he's going to be talking a little bit about the plans that they have for Telepresence for the next few months. It should work.
Richard Lee: Does this work? Okay. Yeah, so thank you for the great demo. So what we're working on. So Telepresence is all open source. We're actually taking it through the CNCF sandbox process right now. From a sort of governance and open sourcing perspective, from a development perspective we're focused a lot of robustness and speed. One of our used case was to talk about is the [inaudible 00:23:56] case.
Richard Lee: So if you've been to Silicone Valley, a lot of developers take [inaudible 00:24:01] from South Bay to up into the valley, and they're bad spot for wi-fi and so you can't do reconnect with Telepresence right now. So we want to support things like reconnect and better performance on behavior when you have bad network connections, improving startup speed. We've done a lot of work around cleaning up. You saw from the demo how we actually swap out the deployment, put in a proxy, and there are certain situations where we would leave that proxy hanging out in the cluster. We want that behavior to improve so we've done a lot of work on our robustness.
Richard Lee: We're doing more deployments at larger organizations that have a lot of our back controls in place. So we've been going through the auditing permissions that Telepresence needs so you need it actually use it. We're also looking at transitioning away from inject TCP because it requires hacking LD Preload or DenLib on your laptop which is actually very difficult from a maintenance perspective. So we're actually exploring other strategies so you can actually support running Telepresence even though you have a VPN client. And then we're also looking at some more future looking used cases around. Things like traffic shadowing, copping production traffic or a percentage of production traffic to your local laptop for production testing, simplifying the UX so you can actually have a config file for configuring it.
Richard Lee: A lot of people are asking us around, "How do I configure this with built system X?" Like they want to run it automatically [inaudible 00:25:28] or use it with particular IDEs and how do you copy environment variables from my local process into my intellogy IDE, so we're trying to do a little bit of work around that. And also supporting [Istios 00:25:40]. So some users have actually manually integrated Telepresence with Istio so that you can actually authenticate to your main website. You know, foo.com, and through the magic of Isteo routing rules actually route those requests straight down to your laptop so you can actually do a real-time development versus your production system.
Richard Lee: So those are some of the things that we're looking at, and we're also really interested in just general feedback around used cases and how people like to use Telepresence. Because, it's very dependent on your developer work flow and we find that everyone has their own sort of way of doing things, so...
Ara Pulido: Cool. Thanks, thank you.
Richard Lee: Thank you.
Ara Pulido: If anyone has any questions? Yes.
Speaker 4: Hi, hello. So perhaps how to go about allowing multiple developers to develop against the same set of backend services running on our cluster? So that they could work concurrently.
Ara Pulido: I don't think there is a way right now. That proxy, so the Telepresence client would talk to a proxy, one of the clients at the proxy. So it won't allow to talk to several ones. Maybe Richard...
Richard Lee: Sorry, yeah so the, so if you can't do it right now so hopefully what we will do is they can bring their own namespaces or clusters for each developer. We are working on some other techniques that let you at least detect when you actually have collision when you drop on the same namespace. But yes, it's a great question. [crosstalk 00:27:34]
Ara Pulido: So I'm gonna repeat the answer for the recording so there is no way right now to do that. The workaround people do is to put different namespaces per developer so you can run the rest of the services together.
Speaker 5: Just to make sure I understand correctly, the only usage currently is for static files changes, or does it include like applications that are developed, let's say with a compiler or something? Like Java based applications or something like that. I mean your demo just covered a static case for changing some HTML but what happens for example for when I [crosstalk 00:28:21] Java based application.
Ara Pulido: Yeah, you can change your Java script so it's not just for static files. So it's basically anything that you're running on your local process would be the one that's talking to the proxy. So it could be any application. It could be a go application.
Speaker 5: Okay, thanks.
Speaker 6: So I was curious, how are you authenticating to the cluster and what permissions you use in order to run this Telepresence?
Ara Pulido: Basically using your Kube CFG. So your Kube Config.
Speaker 7: Hello. First of all, what is the practical use to stream the files from your local mission to the [inaudible 00:29:26] mission?
Ara Pulido: I think that would be a better question maybe for Richard who know more about the internals.
Richard Lee: So we're just using essentially SSHFS for your file system, and we're using, I forget what it's called, but it's an SSH-based VPN for the VPN method. So it's all sort of layer four networking type stuff. We're basically doing a layer four proxy between the cluster and your laptop.
Speaker 7: My second question is: on the solutions for proxying the traffic, have you talked about using [ssx 00:30:05] proxy? That would allow delegating DNS resolution of the local application to the remote cluster.
Richard Lee: Yes, that's what we do. We actually use a fork of [torsocks 00:30:15].
Speaker 8: When you create the connection to your laptop and you have your process running, and then you close you laptop and go home for the day does the Telepresence quit and revert back to the original process?
Ara Pulido: Yeah, basically it closes the connection and goes back to- it's one of what he said they're working on. On actually having being able to reconnect. Yeah, basically goes back to how it was.
Richard Lee: Yeah so when the connection's dropped, when you terminate the Telepresence proxy on your laptop it actually will clean up itself in the cluster. However, if you go to sleep the behavior isn't super nice and that's one of the things we're actually working on. And sort of where architecturally exploring the idea of a persistent proxy on the cluster that you can actually reconnect to. Because right now we deploy the proxy every time you invoke the command.
Speaker 9: So how many developers are working simultaneously at Bitnami on the same project on the same cluster right now and how happy they are with Telepresence can you?
Ara Pulido: Can you repeat the question sir?
Speaker 9: How many developers are currently working on the same project on the same cluster using Telepresence in your project right now?
Ara Pulido: So right now it's mostly three people, but some people are using different clusters. And some of them are using MiniKube as well during the development, so we're not using concurring access of the same cluster.
Speaker 10: It's easy to see how you could fall into some development practices that are unfortunate with the way you're editing your cold starts to diverge from what it actually goes into the images that are built. Can you share something about your experiences and best practices there?
Ara Pulido: Yeah, so actually that one would think that ideally you would move to is the container native, where you build processes is you building an image. The docker image. Then you use the docker method with Telepresence. It allows you to run that new image and you will still be able to make changes, you can still do tricks for faster development like mounting your development folder into that image while you run it. So you can pass any docker and run parameters that you want. So you can still do tricks to go faster but at the same time you make sure that you're building an image and it's similar as possible when you were running production.
Speaker 11: Have you identified any used cases for Telepresence for more than just development? Maybe for CI or for creating a sandbox environment for-.
Ara Pulido: So right now we are just using it for this project. We just started using like two months ago. We haven't been thinking about other use cases. (silence) There is a question over there.
Speaker 6: Sorry, I have another question. So it works fine if you have one instance of your container. What happens if you've got 10 instances of your container in your cluster and you're trying to Telepresence from your machine?
Ara Pulido: Yeah, so what do you mean? Like you have several deployments or?
Speaker 6: So you've got one deployment, several pods right, so you've got 10 pods then your Telepresence, you use Telepresence.
Ara Pulido: So I tried the case where your pod is multi-container, I haven't tried that one but I tried the one that it has one pod, several containers and as far as my testing went, it replaced only the first one. So I don't know what happens if you have several pods. Haven't tried that.
Richard Lee: Yeah, I'm not sure either. Was that your situation?
Ara Pulido: I think, yeah, I think we've run out of time. So, if you have any other questions I'll be at the Bitnami booth most of the conference and I know Richard is going to be at the Datawire one most of the conference. So come reach us. Thank you.